ARTICLE

Recent increases in tropical cyclone intensificationratesKieran T. Bhatia1,2, Gabriel A. Vecchi 2,3, Thomas R. Knutson 1, Hiroyuki Murakami1,4, James Kossin5,

Keith W. Dixon1 & Carolyn E. Whitlock1,6

Tropical cyclones that rapidly intensify are typically associated with the highest forecast

errors and cause a disproportionate amount of human and financial losses. Therefore, it is

crucial to understand if, and why, there are observed upward trends in tropical cyclone

intensification rates. Here, we utilize two observational datasets to calculate 24-hour wind

speed changes over the period 1982–2009. We compare the observed trends to natural

variability in bias-corrected, high-resolution, global coupled model experiments that accu-

rately simulate the climatological distribution of tropical cyclone intensification. Both

observed datasets show significant increases in tropical cyclone intensification rates in the

Atlantic basin that are highly unusual compared to model-based estimates of internal climate

variations. Our results suggest a detectable increase of Atlantic intensification rates with a

positive contribution from anthropogenic forcing and reveal a need for more reliable data

before detecting a robust trend at the global scale.

Corrected: Author correction

https://doi.org/10.1038/s41467-019-08471-z OPEN

1 NOAA/Geophysical Fluid Dynamics Laboratory, Princeton, NJ 08540, USA. 2 Geosciences Department, Princeton University, Princeton, NJ 08544, USA.3 Princeton Environmental Institute, Princeton University, Princeton, NJ 08544, USA. 4University Corporation for Atmospheric Research, Boulder, CO 80307,USA. 5NOAA/National Centers for Environmental Information, Center for Weather and Climate, University of Wisconsin, Madison, WI 53706, USA.6 Engility Inc., Dover, NJ 07806, USA. Correspondence and requests for materials should be addressed to K.T.B. (email: kbhatia@princeton.edu)

NATURE COMMUNICATIONS | (2019) 10:635 | https://doi.org/10.1038/s41467-019-08471-z | www.nature.com/naturecommunications 1

1234

5678

90():,;

Society is faced with significant challenges when tropicalcyclones (TCs) rapidly intensify. These storms quickly risethrough the Saffir–Simpson intensity scale1,2, occasionally

jumping from a category 1 (64–82 knots) to category 5 (>137knots) hurricane within a couple of days. Storms that undergorapid intensification (RI; defined as the 95th percentile of 24-hintensity changes3) reach major hurricane status (wind speedsgreater than 95 knots, categories 3–5 on the Saffir–Simpson Scale)about 80% of the time4 and are associated with the highestforecast errors. As a result, RI can lead to disastrous scenarioswhen coastal areas are not given adequate notice to evacuate andprepare for an extremely intense TC (e.g., Hurricane Audrey,1957)5.

The destruction and forecast error associated with TCs thatundergo RI has inspired new research on whether RI frequencywill be altered due to climate change. A recent statistical-dynamical downscaling study found that the number of TCs thatundergo RI before U.S. landfall is projected to significantlyincrease in the late 21st century compared to the late 20th cen-tury5. A separate study involving climate change simulationsproduced by a high-resolution global climate model (GCM)projected a dramatic increase in the global incidence of RI due toglobal warming6. The agreement in these two studies with dif-ferent methodologies as well as the theoretical backing by a thirdstudy7 suggests we should investigate whether climate change hasalready increased the probability of TCs rapidly intensifying. Ananthropogenically forced signal in recent observational datawould provide evidence that these model projections are well-grounded and therefore, substantial coastal mitigation andadaptation strategies would be required in the near future.

Although TC frequency has stayed approximately constantover recent decades, there is growing evidence that the proportionof TCs which become major hurricanes has significantlyincreased8–13. Therefore, intensification metrics that capture therelative, rather than the absolute, frequency of the highestintensification rates are useful for understanding whether theprobability of RI events is increasing. For example, a recent studyshowed the 95th percentile of 24-h intensity changes significantlyincreased in the central and eastern tropical Atlantic basin duringthe period of 1986–201514. Another intensification metric that isnot dependent on TC frequency, the intensification rate ofintensifying storms, exhibited significant growth between 1977and 2013 in the West Pacific basin15. In both basins, the studiesshowed the large-scale environment becoming more conducive toTC intensification with time. Specifically, areas with the largestincrease in sea surface temperatures (SSTs) and potential inten-sities16 appear to be collocated with the largest positive changes inintensification rates.

Here, we build on these results by evaluating Atlantic andglobal TC intensification trends using different metrics andadditional observational datasets. We focus primarily on theAtlantic basin because it is the basin with the most consistenthigh-quality observations. There are currently no publishedfindings on global trends in TC intensity changes, so these resultsare also included to provide a more comprehensive perspective.To separate anthropogenic trends and normal climate fluctua-tions, we use a state-of-the-art coupled GCM, the high-resolutionforecast-oriented low ocean resolution model (HiFLOR). Speci-fically, we simulate the year-to-year variations in TC intensifi-cation from a suite of HiFLOR multicentury experiments thatserve as a proxy for natural climate variability and compare themagnitude of the trends in the simulated climate to those inobservations . The recent increase in high-intensification rates inthe Atlantic basin is outside the range of normal climate varia-bility defined by HiFLOR, which suggests anthropogenic forcingincreases the likelihood of TCs rapidly intensifying.

ResultsObservational trends. Thus far, TC intensification trend analysishas exclusively focused on International Best-Track Archive forClimate Stewardship17 (IBTrACS) data (Methods), which likelyadds systematic biases to published results. IBTrACS is composedof TC best-track data from different operational agencies whoseTC intensity observations became more reliable when globalsatellite coverage was introduced in the early 1980s (except for theIndian Ocean which did not gain continuous satellite coverageuntil 1998)18. Over much of the globe, satellite data quality hasimproved through increasing resolution and decreasing viewangles, which enhances the accuracy of the most recent intensityestimates. The recent addition of more in situ measurements inthe Atlantic and the East Pacific leads to lower errors in thesebasins, which results in additional spatial and temporal inho-mogeneities in the observations19.

In order to gauge the impact of these inhomogeneities on TCintensification trends, we first compare the probability of 24-hwind speed changes in IBTrACS and a more homogeneous recordof TC intensity, the Advanced Dvorak Technique-HurricaneSatellite-B1 (ADT-HURSAT)12 (Methods). Kossin et al.12 devel-oped ADT-HURSAT in order to maintain the same protocol todetermine TC intensities throughout all ocean basins during theperiod 1982–2009. ADT-HURSAT incorporates artificiallydegraded intensity estimates in certain regions so its data qualityis consistent both spatially and temporally, which is better suitedfor trend analysis. However, this increased data consistencymeans the best technology and analysis techniques are notutilized in ADT-HURSAT, which leads to individual intensityestimates with higher average errors than those in IBTrACS12,20.However, for the scope of our study, which involves assessingrates of change throughout a time series, temporal consistency ispreferable to accuracy in individual estimates.

Figure 1 shows the empirical probability density plotscalculated from ADT-HURSAT and IBTrACS 24-h intensitychanges, using a logarithm transformation on the probabilities tobetter illustrate differences in the tails of the distributions.Supplementary Figure 1 shows the more traditional probabilitydensity plots. For all of our analyses, we impose criteria involvingstorm longevity, location, and intensity to ensure only intensitychanges derived from well-defined storms are included in ourresults (Methods). For the global results, we entirely exclude thenorth and south Indian Oceans from the sample because of thewell-known absence of quality satellite data in the region until199812. Nevertheless, our conclusions do not significantly changewhen these basins are introduced into the analysis.

As expected, the differences in the ADT-HURSAT andIBTrACS intensity change distributions are larger for the globaldata, while the two datasets show more agreement in the Atlanticbasin. For the global results, TCs in ADT-HURSAT maintaintheir intensity for a larger percentage of cases than in IBTrACSbut also have significantly more instances of intensification ratesbetween 50 and 90 knots (resulting in a “shelf” in this part of thedistribution). IBTrACS curves appear almost linear in both plotswhich imply that probability densities are exponentially dis-tributed. Thus, IBTrACS suggests that TC intensification anddissipation are likely controlled by statistically random environ-mental and internal processes21. Meanwhile, ADT-HURSAT’sshelf-like pattern suggests that RI is a special process that couldpreferentially augment the probability of the highest intensifica-tion rates. The former interpretation is much more likelyconsidering the ADT-HURSAT algorithm exhibits well-documented problems during TC eye formation (Methods).

As discussed in Bhatia et al.6, the discrepancies among basinsand datasets cannot be attributed to specific physical processes ordata deficiencies. A summary of the potential explanations for

ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08471-z

2 NATURE COMMUNICATIONS | (2019) 10:635 | https://doi.org/10.1038/s41467-019-08471-z | www.nature.com/naturecommunications

this behavior are included in the Methods section. We here arguethat the agreement among the observational datasets in theAtlantic basin provides some level of confidence in the trendanalysis in this basin. On the other hand, both observationaldatasets have larger uncertainties in global intensity changes.Therefore, any conclusions stemming from the global trendanalysis must be treated with caution.

Figure 2 shows Atlantic basin and global trends in 24-hintensity changes between 1982 and 2009. The 28-year period isselected for analysis because it is the longest continuous samplewhere ADT-HURSAT produces relatively reliable global intensitydata. Quantile regression is calculated for every 5% quantilebetween 5% and 95% (Methods). IBTrACS shows significanttrends in all quantiles except the ones between 35% and 45%.The 95th percentile shows the largest trend of about +4 knotsdecade−1. The negative slopes for the lower percentiles andpositive slopes for the upper quantiles suggests a broadening ofthe intensity change distribution, implying less TCs maintaining asteady intensity and more TCs exhibiting high-intensity fluctua-tions. ADT-HURSAT shares a similar pattern for the slopes of the

quantiles but the trends are more muted. Despite being relativelysmall in magnitude, ADT-HURSAT quantile slopes greater thanthe 50% quantile are statistically significant and positive, with the90th percentile having the largest slope of approximately 1 knotdecade−1.

For the global data, ADT-HURSAT trends are clearly muchsmaller than those in IBTrACS, yet IBTrACS trends are likelymore susceptible to data homogeneity problems. Therefore, it isour interpretation that the weaker global trends shown by ADT-HURSAT are more likely to be correct. This conclusion isparticularly notable because multiple studies4,13–15 have assumedthat IBTrACS is reliable during the time frame considered in ourstudy when in reality, there is likely a spurious trend masked bythe data. It is possible that resolution and algorithm issues withADT-HURSAT prevent it from resolving the intensity changesand have slightly dampened largest trends. However, it is muchmore likely that the influx of new satellites has enabled betterdetection of intensity changes in the later years of the time seriesand thus, artificially enhanced the slope of the extreme quantilesin IBTrACS.

In the Atlantic basin, the quantile regression for ADT-HURSAT and IBTrACS yields similar results for extremeintensity changes. For both observational datasets, the 95thpercentile shows the largest trend. This quantile increases by over+4 knots decade−1 in ADT-HURSAT and +3 knots decade−1 inIBTrACS. The upward trend in TC activity during the 1990s istypically attributed to the Atlantic Multidecadal Oscillation(AMO), which features multidecadal fluctuations in SSTs in theAtlantic basin22. In the late 1990s, the phase of the AMO shiftedfrom negative to positive which coincides with warming in partsof the Atlantic basin where TCs develop14. Therefore, the AMOphase favored RI at the end of the time series. In addition to theAMO behavior, the reliability of the observational data in theAtlantic basin suggests the positive trend in intensification ratesduring 1982–2009 is likely robust. IBTrACS has its lowest errorsin the Atlantic basin because of the plentiful in situ observationsthat are incorporated into the best-track analysis. ADT-HURSATrelies on best-track positions to accurately assign a scene type andfind an eye, which suggests that the additional reconnaissance inthis basin could also help with ADT-HURSAT’s intensityestimates.

As a complement to quantile regression, we also analyze howthe annual rapid intensification ratio (RI ratio) has changed from1982 to 2009. RI ratio is defined as the number of 24-h intensitychanges greater than 30 knots divided by the total number ofintensity changes6, and it reflects changes in the probability of thehighest TC intensification rates. Figure 3 illustrates the Atlanticbasin and global trends in RI ratio for ADT-HURSAT andIBTrACS. For both observational datasets, slopes are positive andsignificant in the global and Atlantic basin data. Much like Fig. 2,the Atlantic basin trends for the observational datasets are verysimilar. Over the 28-year period, the percentage of 24-h intensitychanges that exceed an intensification rate of 30 knotsapproximately triples in IBTrACS and ADT-HURSAT. IBTrACSshows a similar rise in RI ratio for the global data but ADT-HURSAT has a much smaller slope.

Supplementary Figure 2 shows the annual RI ratio of the twoobservational datasets plotted against each other. In the Atlanticbasin, 35.7% of the variance in IBTrACS RI ratio is explained byADT-HURSAT RI ratio. For the global data, the percentageexplained drops to 1.8%, further highlighting the lack ofagreement between the two observational datasets at globalscales. The variations in the slope of RI ratio among the globalobservations as well as the lack of a published theory23 thatexplains the source of a significant positive slope in global TCintensification metrics highlight the uncertainty in the global

log 1

0 (P

)

Global

Atlantic

0

–0.5

–1

–1.5

–2

–2.5

–3

–3.5

–4

–4.5

–5

log 1

0 (P

)

0

–0.5

–1

–1.5

–2

–2.5

–3

–3.5

–4

–4.5

–5

24-h wind speed change (kts)

IBTrACSADT-HURSAT

–110 –90 –70 –50 –30 –10 10 30 50 70 90 110

24-h wind speed change (kts)–110 –90 –70 –50 –30 –10 10 30 50 70 90 110

IBTrACSADT-HURSAT

a

b

Fig. 1 Observed 24-intensity change probability densities. a, b Commonlogarithm of the probability densities calculated from IBTrACS (black) andADT-HURSAT (blue) 24-h intensity changes. a Global and b Atlantic basinresults for the period 1982–2009 are plotted. Data are binned in 20 knotincrements between −110 and 110 knots. The dashed lines indicate the90% confidence interval (between 5th and 95th percentiles of the data)(Methods). All distributions are bounded below by 10−5

NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08471-z ARTICLE

NATURE COMMUNICATIONS | (2019) 10:635 | https://doi.org/10.1038/s41467-019-08471-z | www.nature.com/naturecommunications 3

results. Conversely, the recent uptick in intense TCs and RImagnitude in the Atlantic basin appears more plausible becauseof the AMO shifting phase.

Ordinary trend analysis over a relatively short period(1982–2009) may mistake multidecadal natural variability for along-term trend. For a more refined estimate of the possibleinfluence of natural variability on observed trends and toquantitatively assess whether the recent increase in Atlantic RIratio can be explained by natural variability alone, we comparedthe observed trends to those in a TC-permitting GCM that cansimulate multidecadal climate variability in the Atlantic basin.Global results are also presented with the caveat that the ADT-HURSAT data are likely more realistic because of the homo-geneous time series.

Climate model simulations. HiFLOR is a high-resolution cou-pled GCM that can recover many aspects of the highest TCintensification rates observed in nature and capture the connec-tion between low-frequency climate oscillations and TCbehavior6,24–26. A recent study24 showed that HiFLOR, whennudged toward observed SSTs, can skillfully reproduce theobserved year-by-year variations of the frequency of Category 4and 5 hurricanes (maximum wind speed ≥ 113 kts) in the Atlanticbasin (r ≈ 0.64) and other basins to a lesser extent. HiFLOR’sseasonal predictions of major hurricanes were also skillful in theAtlantic basin, and the high positive correlations betweenHiFLOR predictions and observations are still among the highestdocumented for a dynamical climate model25. The confirmedpresence of important modes of internal climate variability within

HiFLOR as well as the resolved relationship between TCs and thelarge-scale climate instills some confidence that HiFLOR candiagnose whether observed changes are unusual compared toexpected natural variability.

For an initial exploration of whether anthropogenic climatemay be contributing to the increase in TC intensification, wefollow the methodology of Murakami et al.27 and examine TCvariations from a suite of HiFLOR multicentury experiments withdifferent levels of anthropogenic forcing. Four HiFLOR experi-ments are run using anthropogenic forcing (e.g., CO2, aerosols,and ozone) and natural forcing (e.g., volcanic aerosol loading andsolar insolation) representative of the years 1860, 1940, 1990, and2015 (1860CTL, 1940CTL, 1990CTL, and 2015CTL; Methods).

To compare the control simulations to observations, Supple-mentary Figure 3 shows the logarithm of the 24-h intensitychange probabilities observed between 1982 and 1998 inIBTrACS and ADT-HURSAT and those simulated in the1990CTL. As a technique to account for the typical errorassociated with measuring TC intensity, the probability of eachbin was calculated using 1000 subsamples that incorporatedrandom error (Methods). To verify the quality of the controlsimulations, the 17-year period between 1982 and 1998 wasselected for analysis because it is the longest available period withobservational data that is centered on a control simulation. The1990CTL data matches the shape of the IBTrACS curve but thedistribution is not as broad as either of the observational datasets,indicating that HiFLOR underestimates the frequency of rapidweakening and intensification. A discrepancy in the probability ofthe highest intensity changes was also documented by Bhatiaet al.6 but the differences appear larger in Supplementary Figure 3,

IBTrACS Global

IBTrACS Atlantic

ADT-HURSAT Global

ADT-HURSAT Atlantic

Quantile0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Quantile0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Slo

pe (

kts/

year

)

0.5

0.4

0.3

0.2

0.1

0

–0.1

–0.2

–0.3

Slo

pe (

kts/

year

)

0.5

0.4

0.3

0.2

0.1

0

–0.1

–0.2

–0.3

Slo

pe (

kts/

year

)

0.5

0.4

0.3

0.2

0.1

0

–0.1

–0.2

–0.3

Quantile0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Slo

pe (

kts/

year

)

0.5

0.4

0.3

0.2

0.1

0

–0.1

–0.2

–0.3

Quantile0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

a b

c d

Fig. 2 Quantile regression of 24-h intensity changes. a–d Slope of the quantiles for 24-h intensity changes during the period 1982–2009. Slopes are shownfor IBTrACS (a, c) and ADT-HURSAT (b, d) globally (a, b) and in the Atlantic basin (c, d). The black dots represent the slope derived from least squaresregression of intensity change as a function of year for each quantile from 0.05 to 0.95 in steps of 0.05. Shading represents the 5th and 95th percentiles ofthe regressions with randomly perturbed observational data (Methods). The red solid line shows the (constant value) trend in the mean as measured byordinary least squares regression, and the red dotted lines show the 90% confidence interval

ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08471-z

4 NATURE COMMUNICATIONS | (2019) 10:635 | https://doi.org/10.1038/s41467-019-08471-z | www.nature.com/naturecommunications

possibly due to the lack of SST nudging or the different yearssampled.

HiFLOR underestimates the climatological rates of intensifica-tion, which will impact estimates of the amplitude of internalvariability from this model. In order to mitigate the effects ofsystematic biases in HiFLOR intensity changes on estimating theinternal variability of intensification rates, we utilize a statisticaldownscaling technique known as quantile mapping28. Specifi-cally, we test two methods: quantile delta mapping29 (QDM) andbias correction quantile mapping30 (BCQM) (Methods). Allsubsequent analysis only incorporates the QDM algorithmbecause it is more adept at preserving relative changes indistributions of meteorological variables from different climatechange simulations29. Figure 4 shows the effects of both quantilemapping algorithms. In this case, random error is not included in

the probability estimates because we aim to nudge HiFLOR to thebest guess of the intensity distribution. Clearly, the biascorrections are aligning the tails of the 1990CTL more closelywith the tails of the observations. When we explore the potentialinfluence of anthropogenic forcing on TC intensification rates, itis critical that HiFLOR provides a realistic picture of naturalclimate variability. Bias-corrections increase the probability ofextreme intensity changes in the HiFLOR control simulations,which broadens the intensity change distribution. Thus, the bias-corrections enhance the annual variations in RI ratio and therange of the RI ratio slopes in the HiFLOR simulations, whichtranslates into more stringent statistical tests for establishingsignificant observational trends.

To test for observed RI trends that are outside of natural,internal climate variability, overlapping (not independent)

log 1

0 (P

)

0

–0.5Global

Atlantic

–1

–1.5

–2

–2.5

–3

–3.5

–4

–4.5

–5

log 1

0 (P

)

0

–0.5

–1

–1.5

–2

–2.5

–3

–3.5

–4

–4.5

–5

24-h wind speed change (kts)–110 –90 –70 –50 –30 –10 10 30 50 70 90 110

24-h wind speed change (kts)–110 –90 –70 –50 –30 –10 10 30 50 70 90 110

IBTrACSADT-HURSATHiFLOR ORIGHIFLOR QDMHIFLOR BCQM

IBTrACSADT-HURSATHiFLOR ORIGHiFLOR QDMHiFLOR BCQM

a

b

Fig. 4 The effects of quantile mapping on 24-intensity change probabilitydensities. a, b Common logarithm of the probability densities calculatedfrom IBTrACS (black), ADT-HURSAT (blue), HiFLOR 1990CTL (ORIG),HiFLOR QDM-corrected 1990CTL, and HiFLOR BCQM-corrected 1990CTL24-h intensity changes. a Global and b Atlantic basin results are plotted.QDM and BCQM are quantile mapping designed to bias correct 1990CTLto more closely represent the distribution of IBTrACS. ADT-HURSAT andIBTrACS probabilities are computed using the 17-year period centeredaround 1990 (1982–1998). HiFLOR probability density curves are generatedusing 250 years of intensity changes from the 1990CTL. Data is binned in20 knot increments between −110 and 110 knots, and each bin entry isplotted as a dot on a curve. All distributions are bounded below by 10−5

Global

Atlantic

RI c

ases

/tota

l cas

es

0.15

0.1

0.05

0

Year1985 1990 2000 2005 20101995

Year1985 1990 2000 2005 20101995

0.15

0.1

RI c

ases

/tota

l cas

es

0.05

0

IBTrACSADT-HURSAT

IBTrACSADT-HURSAT Slope = 1.54e–03 ± 1.17e–04

Slope = 1.96e–04 ± 4.90e–05

Slope = 1.40e–03 ± 3.65e–04Slope = 1.24e–03 ± 2.22e–04

a

b

Fig. 3 Rapid intensification ratio trends. a, b Observed trends in the rapidintensification (RI) ratio of ADT-HURSAT (black) and IBTrACS (blue) overthe 28-year period 1982–2009 using a global and b Atlantic data. RI ratio isdefined as the number of 24-h intensity changes above 30 knots divided bythe total number of 24-h intensity changes. Trends in the time series of theannual mean RI ratio are denoted by dashed lines. The slopes of the trendlines as well as their 90% confidence intervals are provided. The slopes andconfidence intervals are calculated using 1000 randomly perturbed samplesof the observational data. Shading represents the 5th and 95th percentilesof the 1000 regressions with these randomly perturbed observational data(Methods)

NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08471-z ARTICLE

NATURE COMMUNICATIONS | (2019) 10:635 | https://doi.org/10.1038/s41467-019-08471-z | www.nature.com/naturecommunications 5

28-year RI ratio slopes in the bias-corrected HiFLOR controlsimulations are compared to the ADT-HURSAT and IBTrACS RIratio slopes between 1982 and 2009. Supplementary Figure 4shows how the probability distribution of RI ratio slopes changeswhen selecting 28-year segments from HiFLOR 1860CTL that areindependent and not independent. The plotted bins show theindependent sample has a slightly broader distribution but ingeneral, the probabilities of the bins are similar to those in theoverlapping bins.

Figure 5 contains a box and whisker plot that shows thedistribution of RI ratio slopes for the QDM-corrected 1860CTL.The observed slopes during 1982–2009 are overlaid on each plot.Supplementary Figure 5 is similar to Fig. 5 but shows the box andwhisker plots for all of the HiFLOR control simulations.Supplementary Figure 6 portrays the corresponding cumulativedistribution functions (CDF) of the RI ratio slopes for all of thecontrol simulations as well as the observed RI ratio slopes inADT-HURSAT and IBTrACS. In the Atlantic basin, the slope ofthe RI ratio for ADT-HURSAT and IBTrACS are both above the99th percentile of the slopes of any of the bias-corrected HiFLORcontrol simulations. Therefore, the large positive slope of RI ratioin both observational datasets is outside HiFLOR’s estimate ofexpected internal climate variability, which suggests the model’sdepiction of climate oscillations like the AMO cannot explain theobserved trend.

For the global data, IBTrACS data indicates a trend that is wellabove the bounds of the observed natural variability of RI ratio inany HiFLOR control simulation, but the noted temporal andspatial heterogeneities leads us to question the validity of thisresult. The slope of RI ratio for ADT-HURSAT is approximatelythe 80th percentile of the HiFLOR 1860CTL, indicating theobserved trend is not significant compared to modelled climatevariability. Importantly, ADT-HURSAT observed data also yieldsan RI slope that is still relatively unusual compared to theHiFLOR preindustrial natural variability. This emerging trend isconsistent with an increasing trend in the global intensity ofTCs12. In the near future, ADT-HURSAT will be extended to2016, and it is therefore vital for future analysis to determinewhether the global trend also becomes significant.

In order to explore whether the observed changes between1982 and 2009 could be attributable to anthropogenic forcing, wecompare RI ratio in the three climate change simulations to thepreindustrial 1860CTL. Figure 6 shows the percent difference inRI ratio between 1860CTL and the HiFLOR simulations withstronger anthropogenic forcing (1940CTL, 1990CTL, and2015CTL). The analysis only involves HiFLOR simulations sobias corrections and random error are not applied to the data(Methods). Grid boxes that are not statistically significant aredemarcated with a white “X”. The prevalence of red boxesthroughout the maps conveys that a larger percentage of TCs areundergoing RI in 1940CTL, 1990CTL, and 2015CTL than in1860CTL. There are no significant blue grid boxes in the1990CTL and 2015CTL maps, and the number of gridboxes thatrecord significant increases in RI ratio is 41 (9.7% of all shadedgridboxes) in 1940CTL, 156 (34.7% of all shaded gridboxes) in1990 CTL, and 164 (38.2% of all shaded gridboxes) in 2015CTL.Historical radiative forcing changes applied to HiFLOR clearlylead to more frequent large intensification rates, but the availablesimulations do not allow us to strictly attribute the identifiedgrowth in RI during the period 1982–2009 to anthropogenicforcing. An ensemble of experiments that simulate anthropo-genically forced climate change during this specific period,including changes in RI ratio, would be better suited for morecomprehensive attribution analysis. Thus, for now, we can onlyconclude that anthropogenic forcing significantly increasesextreme TC intensification rates in the HiFLOR model comparedto preindustrial (1860CTL) conditions.

In the two most reliable long-term observational recordsavailable for TC intensity changes, the proportion of 24-h TCintensification rates greater than 30 knots significantly increasesin the Atlantic basin between 1982 and 2009. Results averagedover all basins show a significant increase in TC intensificationrates in IBTrACS but not in ADT-HURSAT. By itself, a 28-yearupward trend in a TC intensification metric does not necessarilyreflect the effects of anthropogenic climate forcing because of the

2

1

0

–1

–2

Slo

pe o

f RI r

atio

Slo

pe o

f RI r

atio

2

1

0

–1

–2

× 10–3

× 10–3

Global

p values ~ 0.01

p value = 0.20

p value < 0.01

Atlantic

IBTrACS

ADT-HURSAT

IBTrACS

ADT-HURSAT

a

b

Fig. 5 Observed trends in RI ratio vs. 1860CTL natural variability. a, b Boxand whisker plot represents the distribution of slopes of RI ratio in theQDM-corrected 1860 HiFLOR control simulation. a Global and b Atlanticbasin results are plotted. Each slope is calculated by applying least squaresregression analysis to annual RI ratio values in overlapping 28-year periods.Thus, the number of slopes for a control simulation is the number ofavailable years subtracted by 28 (i.e., 1860CTL has 1422 slopes). The redline in each box indicates the median of the slopes. The box is bounded bythe 25th and 75th percentiles of the data, and the whiskers bracketapproximately 99% of the data. Red plus signs indicate outliers whosevalues are outside of whiskers’ range. IBTrACS and ADT-HURSAT trends inannual mean RI ratio between 1982 and 2009 are respectively representedby blue and green dotted lines and the corresponding p values are listedbelow each line

ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08471-z

6 NATURE COMMUNICATIONS | (2019) 10:635 | https://doi.org/10.1038/s41467-019-08471-z | www.nature.com/naturecommunications

intrinsic natural variability of the climate system. Here, we usebias-corrected TC intensification rates simulated by HiFLOR todemonstrate that natural variability cannot explain the magnitudeof the observed upward trend in the Atlantic basin. Theseconclusions are possible because HiFLOR is a unique climatemodel that can successfully simulate the most intense TCs andhighest intensification rates in multicentury simulations.

This study is limited by the ability of a climate model toaccurately represent natural variability as well as the uncertaintyaround the trends in relatively short observational records.However, this study represents a crucial first step in quantifyingthe precise roles of stochastic processes, anthropogenic warming,and natural variability when assessing changes in TC intensifica-tion rates. Further analysis with additional high-resolution

1940CTL

1990CTL

2015CTL

300° W 270° W 240° W 210° W 180° W 150° W 120° W 90° W 60° W 30° W

200

150

100

50

0

–50

–100

–150

–200

Per

cent

diff

eren

ce in

RI c

ases

/tota

l cas

es

200

150

100

50

0

–50

–100

–150

–200

Per

cent

diff

eren

ce in

RI c

ases

/tota

l cas

es

200

150

100

50

0

–50

–100

–150

–200

Per

cent

diff

eren

ce in

RI c

ases

/tota

l cas

es

40° N

20° N

40° S

20° S

0°

40° N

20° N

40° S

20° S

0°

40° N

20° N

40° S

20° S

0°

300° W 270° W 240° W 210° W 180° W 150° W 120° W 90° W 60° W 30° W

300° W 270° W 240° W 210° W 180° W 150° W 120° W 90° W 60° W 30° W

a

b

c

Fig. 6 Anthropogenic forcing’s effects on RI ratio in HiFLOR. a–c Simulated changes in RI ratio by the 1940CTL (a), 1990CTL (b), and 2015CTL (c) relativeto the 1860CTL. Percent difference in RI ratio between HiFLOR 1860CTL and each climate change simulation is plotted in each 5° × 5° grid box. Data is onlyplotted in a grid box if at least one TC passes through the grid box every 50 years in the two experiments used to calculate percent difference. Red (blue)squares indicate grid boxes where a larger (smaller) percentage of 24-h intensity changes exceed 30 knots in the climate change simulations than in the1860CTL. Grid boxes that achieve a p value of 0.05 using a binomial proportion test are considered statistically significant. White “Xs” are located in gridboxes that are not statistically significant

NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08471-z ARTICLE

NATURE COMMUNICATIONS | (2019) 10:635 | https://doi.org/10.1038/s41467-019-08471-z | www.nature.com/naturecommunications 7

climate models and a longer and more reliable observationalrecord is required to confirm these conclusions. Regardless, thesetrends provide another example of the potentially seriousrepercussions of anthropogenic warming for TCs.

MethodsObserved data. We used the International Best Track Archive for ClimateStewardship (IBTrACS)17, v03r09, and the Advance Dvorak Technique-HurricaneSatellite-B112 (ADT-HURSAT) for the period 1982–2009. For our IBTrACS ana-lysis, we only consider best-track data from the National Hurricane Center for theAtlantic and east Pacific and the Joint Typhoon Warning Center for the remainderof the globe17. One of the benefits of only using data from these U.S. agencies isthey follow the same definition of maximum winds: the highest 1-min average at10 m height over a smooth surface31. Best-track data start as operational estimatesof the intensity and track of a TC and are refined at the end of a TC’s lifetime witha combination of in situ (e.g., dropsondes, scatterometers, and buoys), radar, andsatellite measurements. Best-track intensity and position estimates are availableevery 6 h at the four synoptic times (0000, 0600, 1200, and 1800 UTC) and arerecorded to the nearest 5 knots (1 kt= 0.5144 m s−1) and 0.1° latitude/longitude32.

The creation of ADT-HURSAT consists of four main steps. Geostationarysatellite imagery is first analyzed from International Satellite Cloud ClimatologyProject (ISCCP)-B1 data33–35. Then, the data are centered on IBTrACS TCs andsubsampled to be both spatially and temporally homogeneous. Finally, a simplifiedversion of the advanced Dvorak technique20 is used to evaluate the data anddetermine a maximum TC wind speed. ADT-HURSAT data are produced every 3h based on satellite data that has been uniformly subsampled to a horizontalresolution of 8 km, and wind speeds are recorded to the nearest tenth of a Dvorak“T-number” (depending on the current intensity, between 1 and 3 knots).

Although ADT-HURSAT and IBTrACS represent the two most reliable multi-decadal observational datasets for TC intensification trend analysis, they both havelimitations. In IBTrACS, Atlantic and East Pacific in situ observations have rapidlyimproved in the last 30 years, while other basins were rarely able to obtain aircraftreconnaissance. In fact, the west Pacific was the only other basin that had access toaircraft reconnaissance but these observations stopped in 1987. As a result, thenumber of measurements available in different basins varies considerably, leadingto temporal and spatial inconsistencies in IBTrACS observational quality. Betterintensity estimates at the end of the historical record potentially causes heighteneddetection of the larger intensity changes. Additionally, best-track estimates are alsosusceptible to human error because they are assembled at operational centers.

Compared to IBTrACS, ADT-HURSAT employs reduced-resolution satelliteimagery, both in space and time. The omission of in situ measurements toaccompany the coarsened satellite imagery is especially problematic for ADT-HURSAT identifying “scene type” changes when an eye is forming. As a result,there is a well-documented unphysical plateau at the eye formation stage of TCdevelopment in intensity observations from ADT-HURSAT. Low intensity changesare present for long periods of time and then the ADT-HURSAT is forced “to catchup” with higher intensity changes once it finally detects an eye. Thus, peak TCintensities (and changes) may not be fully resolved by ADT-HURSAT, especially inthe cases of a transient, small eye. However, these deficiencies remain consistentthroughout the time series, which suggests that there is a low likelihood of asystematic bias in ADT-HURSAT trends. It is also possible that larger eyes and alonger mean duration of the “plateau” period for TCs in the Atlantic basin couldlead to minimal effects of these deficiencies.

Criteria for inclusion in sample. For consistency, intensity change values inHiFLOR and the observational datasets are rounded to the nearest five knots. Weonly consider TCs that are active for at least 72 h and exceed wind speeds of 34knots for at least 36 h. We restrict our analysis sample to only consider cases wherethe TC center is located over the ocean, the starting and ending TC position arebelow 40° of latitude, and the TC intensity stays above 34 knots. The warm corecriteria discussed in Murakami et al.24 is also applied to the HiFLOR data beforeanalysis.

Control experiments. Four HiFLOR control simulations introduced in Murakamiet al.27 were used here to represent natural climate variability and provide theframework for exploring anthropogenic effects. Control simulations were createdusing anthropogenic forcing fixed at 1860 (1860CTL), 1940 (1940CTL), 1990(1990CTL), and 2015 (2015CTL) levels. Owing to limited computational resources,1860CTL, 1940CTL, 1990CTL, and 2015CTL were run for different lengths: 1500,200, 300, and 200 years, respectively. The first 50 years of all simulations weredisregarded to mitigate effects of model drift. Basic conclusions remained similareven when we using the smallest sample size (150 years) for all the controlsimulations. The fixed forcing agents for the control simulations were atmosphericCO2, CH4, N2O, halons, tropospheric and stratospheric O3, anthropogenic tro-pospheric sulfates, black and organic carbon, and solar irradiance.

Uncertainty quantification. ADT-HURSAT and IBTrACS respectively provideintensity estimates to the nearest 1–3 and 5 knots. We use Monte Carlo techniques

to create random noise before analyzing the discretized data. Random noise pre-vents multiple data points from having the same value and provides an estimate ofthe typical error associated with measuring TC intensity. 1000 subsamples wereproduced by adding random noise from a uniform distribution on the interval±2 ´

ffiffiffiffiffi

50p

knots to each intensity change value. The magnitude of this randomnoise is derived by adding 5 knots of error in quadrature (propagation of errorsstemming from the intensity change calculation), which is a conservative estimatefor the typical error associated with each TC intensity observation35,36. To calculatethe slope of any intensification metric for ADT-HURSAT and IBTrACS, we use theslope of the mean of the 1000 subsamples for that particular metric. For example,quantile regression involved the calculation of every 5th percentile in each of the1000 subsamples for each year. One-thousand slopes of each percentile were cal-culated and the mean of the slopes was considered the best estimate of the1982–2009 slope of the quantile. The 5th and 95th percentiles of the 1000 slopeswere considered the uncertainty bounds for the quantile.

For the creation of Fig. 6, random error is not added to the data because we onlycompare spatial differences in RI ratio among the different HiFLOR simulations.Statistical significance is computed using a binomial proportion test with p valuesbelow 0.05 considered significant37. Data is only plotted in a grid box if there is atleast one-fourth of a TC day per year in the two HiFLOR simulations used tocalculate the percent difference.

Quantile mapping. In this work, two univariate bias correction methods weretested. We primarily focus on results produced using an additive version of QDM29

by making use of R programming language code contained in the CRAN MBCpackage version 0.10–438. As reported in Appendix A of Cannon et al.29, theadditive version of QDM is functionally very similar to the equidistant CDFmatching algorithm of Li et al.39. The formulation of QDM aims to preserverelative changes in model-simulated climate variable quantiles. In other words,with respect to the quantiles, QDM guards against distorting the underlying cli-mate model’s climate sensitivity. For comparison purposes, we also used a simplerand more standard quantile mapping approach called BCQM. The more com-monly used standard BCQM method develops transfer equations solely fromcomparisons of observations and model data sets drawn from a common historical(baseline) period, which can yield results in which the relative trends in biascorrected output differ from those of the raw GCM results. Meanwhile, the QDMmethod also incorporates adjustments accounting for differences between themodeled distributions of the historical period and future projections. As is gen-erally the case for statistical bias correction methods, QDM and BCQM bothassume that the GCM biases present in the historical period are consistent withbiases in the GCM’s future projections (i.e., a stationarity assumption).

For both of these mappings, IBTrACS serves as the training dataset that1990CTL data is nudged to because its observations have lower errors than ADT-HURSAT. ADT-HURSAT was also tested as the target distribution for the biascorrection, but these results are excluded since the basic conclusions of the resultsremain the same. After establishing the transfer function for the bias correctionusing 1990CTL, the other control simulations are also bias-corrected.

Code availability. The code that supports the findings of this study is availablefrom the corresponding author on request.

Data availabilityThe source code of the climate model can be found at https://www.gfdl.noaa.gov/cm2-5-and-flor. The data that support the findings of this study are available fromthe corresponding author on request.

Received: 24 August 2018 Accepted: 8 January 2019

References1. Saffir, H. S. Hurricane wind and storm surge. Mil. Eng. 423, 4–5

(1973).2. Simpson, R. H. The hurricane disaster potential scale. Weather Wise 27,

169–186 (1974).3. Kaplan, J. & DeMaria, M. Large-scale characteristics of rapidly intensifying

tropical cyclones in the North Atlantic basin.Weather Forecast. 18, 1093–1108(2003).

4. Lee, C.-Y., Tippett, M. K., Sobel, A. H. & Camargo, S. J. Rapid intensificationand the bimodal distribution of tropical cyclone intensity. Nat. Commun. 7,10625 (2016).

5. Emanuel, K. Will global warming make hurricane forecasting more difficult?Bull. Am. Meteor. Soc. 98, 495–501 (2017).

6. Bhatia, K., Vecchi, G., Murakami, H., Underwood, S. & Kossin, J. Projectedresponse of tropical cyclone intensity and intensification in a global climatemodel. J. Clim. 31, 8281–8303 (2018).

ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08471-z

8 NATURE COMMUNICATIONS | (2019) 10:635 | https://doi.org/10.1038/s41467-019-08471-z | www.nature.com/naturecommunications

7. Emanuel, K. Self-stratification of tropical cyclone outflow. Part II: implicationsfor storm intensification. J. Atmos. Sci. 69, 988–996 (2012).

8. Emanuel, K. A. Increasing destructiveness of tropical cyclones over the past 30years. Nature 436, 686–688 (2005).

9. Elsner, J. B., Kossin, J. P. & Jagger, Y. H. The increasing intensity of thestrongest tropical cyclones. Nature 455, 92–95 (2008).

10. Holland, G. & Bruyère, C. L. Recent intense hurricane response to globalclimate change. Clim. Dyn. 42, 617–627 (2014).

11. Hoyos, C. D., Agudelo, P. A., Webster, P. J. & Curry, J. A. Deconvolution ofthe factors contributing to the increase in global hurricane intensity. Science312, 94–97 (2006).

12. Kossin, J. P., Olander, T. L. & Knapp, K. R. Trend analysis with a new globalrecord of tropical cyclone intensity. J. Clim. 26, 9960–9976 (2013).

13. Wang, C., Wang, X., Weisberg, R. H. & Black, M. L. Variability of tropicalcyclone rapid intensification in the North Atlantic and its relationship withclimate variations. Clim. Dyn. 49, 3627–3645 (2017).

14. Balaguru, K., Foltz, G. R., & Leung, L. R., Increasing magnitude of hurricanerapid intensification in the central and eastern tropical Atlantic. Geophys. Res.Lett. 45, 4238–4247 (2018).

15. Mei, W. & Xie, S. P. Intensification of landfalling typhoons over the northwestPacific since the late 1970s. Nat. Geosci. 9, 753–757 (2016).

16. Emanuel, K. A. Thermodynamic control of hurricane intensity. Nature 401,665 (1999).

17. Knapp, K. R., Kruk, M. C., Levinson, D. H., Diamond, H. J. & Neuman, C. J. Theinternational best track archive for climate stewardship (IBTrACS): unifyingtropical cyclone best track data. Bull. Am. Meteorol. Soc. 91, 363–376 (2010).

18. Schreck, C. J., Knapp, K. R. & Kossin, J. P. The impact of best trackdiscrepancies on global tropical cyclone climatologies using IBTrACS. Mon.Weather Rev. 142, 3881–3899 (2014).

19. Uhlhorn, E. W. et al. Hurricane surface wind measurements from anoperational stepped frequency microwave radiometer. Mon. Weather Rev.135, 3070–3085 (2007).

20. Olander, T. L. & Velden, C. S. The advanced Dvorak technique: continueddevelopment of an objective scheme to estimate tropical cyclone intensity usinggeostationary infrared satellite imagery. Weather Forecast. 22, 287–298 (2007).

21. Kowch, R. & Emanuel, K. Are special processes at work in the rapidintensification of tropical cyclones? Mon. Weather Rev. 143, 878–882 (2015).

22. Goldenberg, S. B., Landsea, C. W., Mestas-Nuñez, A. M. & Gray, W. M. Therecent increase in Atlantic hurricane activity: causes and implications. Science293, 474–479 (2001).

23. Walsh, K. & Coauthors Hurricanes and climate: the US CLIVAR workinggroup on hurricanes. Bull. Am. Meteorol. Soc. 96, 997–1017 (2015).

24. Murakami, H. et al. Simulation and prediction of category 4 and 5 hurricanesin the high-resolution GFDL HiFLOR coupled climate model. J. Clim. 28,9058–9079 (2015).

25. Zhang, W. et al. Improved simulation of tropical cyclone responses to ENSOin the Western North Pacific in the high-resolution GFDL HiFLOR coupledclimate model. J. Clim. 29, 1391–1415 (2016).

26. Murakami, H. et al. Seasonal forecasts of major hurricanes and landfallingtropical cyclones using a high-resolution GFDL coupled climate model. J.Clim. 29, 7977–7989 (2016).

27. Murakami, H., Gabriel, A. Vecchi & Underwood, Seth Increasing frequency ofextremely severe cyclonic storms over the Arabian Sea. Nat. Clim. Change 7,885–889 (2017).

28. Panofsky, H. A. & Brier, G. W. Some Applications of Statistics to Meteorology,191–207 (Pennysylvania State University Press, University Park, 1958).

29. Cannon, A. J., Sobie, S. R. & Murdock, T. Q. Bias correction of GCMprecipitation by quantile mapping: how well do methods preserve changes inquantiles and extremes? J. Clim. 28, 6938–6959 (2015).

30. Lanzante, J. R., Dixon, K. W., Nath, M. J., Whitlock, C. E. & Adams-Smith, D.Some pitfalls in statistical downscaling of future climate. Bull. Am. Meteor.Soc. 99, 791–803 (2018).

31. Harper, B. A., Kepert, J. & Ginger, J. Guidelines for converting between variouswind averaging periods in tropical cyclone conditions. WMO/TD 1555. 54(World Meteorological Organization, 2010).

32. Landsea, C. W. & Franklin, J. L. Atlantic hurricane database uncertainty andpresentation of a new database format. Mon. Weather Rev. 141, 3576–3592(2013).

33. Knapp, K. R. Calibration of long-term geostationary infrared observationsusing HIRS. J. Atmos. Ocean. Technol. 25, 183–195 (2008).

34. Knapp, K. R. Scientific data stewardship of International Satellite CloudClimatology Project B1 global geostationary observations. J. Appl. RemoteSens. 2, 023548 (2008).

35. Knapp, K. R. & Kossin, J. P. New global tropical cyclone data set from ISCCPB1 geostationary satellite observations. J. Appl. Remote Sens. 1, 013505 (2007).

36. Torn, R. D. & Snyder, C. Uncertainty of tropical cyclone best trackinformation. Weather Forecast. 27, 715–729 (2012).

37. Suissa, S. & Shuster, J. J. Exact unconditional sample sizes for the 2 × 2binomial trial. J. R. Stat. Soc. Ser. A 148, 317–327 (1985).

38. Cannon, A. J., MBC: multivariate bias correction of climate model outputs (Rpackage). https://CRAN.R-project.org/package=MBC (2018).

39. Li, H., Sheffield, J. & Wood, E. F. Bias correction of monthly precipitation andtemperature fields from Intergovernmental Panel on Climate Change AR4models using equidistant quantile matching. J. Geophys. Res. 115, D10101(2010).

AcknowledgementsThe authors thank Dr. Benjamin Bronselaer and Dr. Nathaniel Johnson for their sug-gestions and comments during the internal review process. The authors also would liketo acknowledge Youngrak Cho for his edits to multiple figures. Kieran Bhatia and GabrielVecchi were supported by National Science Foundation under Grant AGS-1262099 andin part by the Carbon Mitigation Initiative at Princeton University BP International02085(7).

Author contributionsK.B. designed the study, completed a majority of the analysis, and wrote the entirety ofthe manuscript. G.V. helped develop the HiFLOR model, and G.V. and H.M carried outthe experiments. G.V. also provided key suggestions on research topics to explore. H.M.provided key suggestions on the analysis. T.K. assisted in the statistical analysis andprovided guidance on the analysis. J.K. developed ADT-HURSAT data and provided itfor analysis. J.K. provided minor suggestions on experimental design. K.D. helped designand write a description of the bias corrections for HiFLOR. C.W. was the primarydeveloper of the bias corrections for HiFLOR.

Additional informationSupplementary Information accompanies this paper at https://doi.org/10.1038/s41467-019-08471-z.

Competing interests: The authors declare no competing interests.

Reprints and permission information is available online at http://npg.nature.com/reprintsandpermissions/

Journal peer review information: Nature Communications thanks the anonymousreviewers for their contribution to the peer review of this work. Peer reviewer reports areavailable.

Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims inpublished maps and institutional affiliations.

Open Access This article is licensed under a Creative CommonsAttribution 4.0 International License, which permits use, sharing,

adaptation, distribution and reproduction in any medium or format, as long as you giveappropriate credit to the original author(s) and the source, provide a link to the CreativeCommons license, and indicate if changes were made. The images or other third partymaterial in this article are included in the article’s Creative Commons license, unlessindicated otherwise in a credit line to the material. If material is not included in thearticle’s Creative Commons license and your intended use is not permitted by statutoryregulation or exceeds the permitted use, you will need to obtain permission directly fromthe copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.

© The Author(s) 2019

NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08471-z ARTICLE

NATURE COMMUNICATIONS | (2019) 10:635 | https://doi.org/10.1038/s41467-019-08471-z | www.nature.com/naturecommunications 9